Manuel DeLanda...Meshwork or Hierarchy?...Doors 2
< < < Previous Page | Speakers | Conference Report | Doors 2 | Mediamatic Index
The recent development of theories of nonlinear dynamics and of processes of self-organization has given these critics a boost. While before the 1960's it was virtually imposible to imagine the emergence of order without a central agency behind it, today we are familiar with a growing body of knowledge about the spontaneous generation of ordered structures in inorganic as well as organic (and even social) processes. For the purposes of understanding the issue of home territories, it will be useful to trace the effect of these new ideas in the current confrontation between symbolic Artificial Intelligence (which retains a hierarchical organization of centers) and the new connectionist school, based on nonlinear dynamics and a decentralized conception of the mind. An `artificial bird's brain' designed with symbolic AI would typically contain representations of the world (coded in birds mental language) forming a cognitive map of the animal's surroundings. Creating a territory would then consist in symbolic operations performed on these representations and only later implemented as actions in the real world. A connectionist approach, on the other hand, would be to generate a population of neural nets, each of which is dynamically connected to the outside world. In other words, without using mental representations each neural net in the bird's brain is in a nonlinear stable state (or attractor) which is associated with a similarly stable pattern in the animal's enviroment. A pattern outside (such as the expresive qualities of a territorial marker) can then be recognized by the animal without forming an explicit internal symbol to stand in for the pattern.
Neural nets have indeed supplied us with a concrete technological paradign of how brains could function without internal homunculi. Unlike symbolic AI who has only scored successes in the modelling of evolutionarly late skills (such as playing chess or proving theorems), connectionist designs have succeded in capturing some more basic abilities, such as face recognition. And yet, for our purposes here, not even this novel branch of cognitive science has gone far enough. The real breakthrough to understand how home territories could self-organize through brains and outside expressive qualities, comes from an even younger branch of AI: behavioral-based AI (or as it is sometimes called, the animat approach). The differences between behavioral and symbolic AI have been very lucidly expressed by Pattie Maes, and we may summarize them as follows: Symbolic AI decomposes minds into relatively large functional modules (perception, execution) interfaced together by central representations (beliefs, desires, intentions). The activity of the modules and the representations form a static `model of the world', and the effects of learning are conceived as the operation of reformulating this model. Behavioural AI, on the other hand, does not involve high level general modules (which as I said, almost always embody homunculi) but low level specific modules (such as `collision avoidance'). High level skills emerge out of the interactions of these micro-modules, none of which can be said to possess the skill. More importantly for our present purposes, behavioural AI does not aim at the internal generation of a world model, but rather, it situates its robotic animals in the real world so that the objective features of the enviroment can be used as a form of external memory. This modelling strategy is sometimes expressed with the phrase: `The world is its own best model'.
One useful way of explaining this rather cryptic phrase is by using some insights from the ecological theory of perception developed by James Gibson in the 1960's. Gibson elaborated the crucial idea that the enviroment provides an animal with meaningful contraints which he called `affordances'. For instance, solid ground supplies animals with (or `affords' them) a surface to walk on. On reaching the edge of a swamp an animal's `muscular intelligence' tells it automatically that the ground there does not afford suitable support, and the animal reaches this `conclusion' without the need for an internal `world model' which includes representations of dry and wet land. Similarly, a hole in the ground of suitable size affords an escaping animal a place to hide, and twigs afford the bird nest-construction materials. An open enviroment affords locomotion in all directions, while a cluttered one affords it only at certain openings. And, of course, what a given part of the world affords depends on the animal: water, due to surface tension, affords a walking surface to a small insect but not to a large bird, to whom it affords at most a gliding surface. The point of all this is that the world possess a kind of intrinsic `proto-semantics', which are meaningful to animal minds in a functional way.
In terms of behavioural AI this means that, a simple module for collision-avoidance (so simple it does not contain a homunculus) together with the obstacles afforded by a rooms's walls can generate the complex behaviour of `wall following' without an internal representation of the room. But the layout of surfaces in the enviroment is only one source of affordances, the behaviour of other animals is too. Prey afford predators nutrition, while a territorial bird affords another competition. Animals may also afford one another opportunities for cooperation. This idea has also been exploited by behavioral AI in designs where novel intelligent behaviors emerge not only from the interaction animal-enviroment, but also from the interactions between the animals themselves. Hence the idea of building not expensive single robots, but teams of relatively inexpensive ones. This has the advantage that the solution to a given problem emerges out of the interactions of the whole team, with no single member being essential to the task. In this way, the inevitable breakdowns and malfunctions that plague any real life applications do not cripple the entire enterprise, as would be the case with the single robot approach.
By now it should be clear what I am getting at. Home territories self-organize through a complex interplay between male and female birds and the expressive affordances of their enviroment. For example, the male satin bowerbird builds a stage decorated with bright blue objects of different kinds with which he tempts a female to stop by. Then, as the courtship begins, he will grab a yellow flower in his beak and alternately display it and hide it in a species-specific ritual. The home territory of the couple may be seen as emerging from simple in-the-head components (which are partly learned, partly inherited) and the optical affordances of the blue and yellow objects. Now, is it possible to extend these remarks to human beings? Is it possible that our own homes self-organize in this way, with the expressive affordances of our cherished possessions playing an active role too?
Although I would like to answer this question affirmatively, there are other aspects of the problem that we must consider first. In particular, unlike birds we possess also linguistic abilities, and hence a greater propensity to form representations and plans inside our heads. It may be, as philosopher Andy Clark has suggested, that our minds are a kludge (or bricollage) of different kinds of intelligence: some intelligent abilities arise out of decentralized and parallel processes, others from centralized and sequential ones. One useful way to think about this is to view the evolution of the human mind as involving a similar process as symbolic AI, only in reverse. Let me explain. When the first AI programs were written, programming languages and computer hardware were very hierarchical and sequential. In the 1970's when symbolic AI switched to the creation of expert-systems, the need for flexibility forced them to create programing languages which simulated parallel processing even while running in sequential hardware. Andy Clark's idea is that our evolution may have involved a similar, though opposite, solution: we began with a highly parallel and non-hierarchical hardware (like birds) and at some point our brains began to simulate a sequential and centralized mind: the stream of lingusitic consciousnes with which we are familiar through introspection.
If our minds are thus hybrids of two or more computer-types then we should expect our homes to be also complex mixtures of self-organized and planned components, or to use the technical terms, of hierarchies and meshworks. Hierarchies are structures in which components have been sorted out into homogenous groups, then articulated together. Meshworks, on the other hand, articulate heterogenous components as such, without homogenizing. A bird's territory is more meshwork than hierarchy, while the hypothetical pre-furnished corporate apartment I mentioned at the beggining of this paper, has more hierarchy than meshwork elements in it. Our homes can then be seen as mixtures of self-organized and planned components: certain objects will occupy a space and fulfill a function which we deliberatedly assigned to them while others will be located where they meshed well with their surroundings. And in these terms, the feeling of home could be derived from how well we mesh with the objects and expressive affordances of this private enviroment.
The concepts of meshwork and hierarchy have become one of the cornerstones for the application of nonlinear dynamical simulations to social and economic questions. Hence they are very useful in analysing not only the structure of our private spaces, but also that of public spaces. That is, they help us thinking not only about our homes but also about the home of our homes: the city. From this point of view our individual homes become households, one of several types of institutions housed by our home towns. These institutional populations are also complex mixtures of meshworks and hierarchies, of markets and bureocracies, for example. Pre-capitalist markets, like those which existed in medieval Europe, in China or India, or indeed in many small towns even today, are structures that emerge out of a decentralized decision-making process which brings heterogenous needs and offerings together. In modern nonlinear models, markets have very little to do with the `invisible hand', involving complex processes of self-organization and not just demand and supply. Behavioural AI (as well as other forms of nonlinear cognitive science) sometimes use market-like structures (such as bidding schemes) to replace centralized decision-making in the robot's mind.
Speakers | Conference Report | Doors 2 | Mediamatic Index | Next Page > > >
Last updated: 16 feb1995
